Video Stabilization Quality Assessment In The Space Of Lie Group

In recent years,there has been a multitude of successful algorithms developed for video stabilization.However,how to objectively and accurately assess the effectiveness of these algorithms is a problem that has not been thoroughly explored.Due to the lack of the appropriate evaluation mechanism to quantitatively measure the stabilization of video motion,it leads to the lack of objective evaluation of the stabilization algorithm performance,and also makes users unable to choose the video stabilization algorithm to meet the demand.Similar to image quality assessment,video stabilization quality assessment needs an appropriate space and chooses an appropriate measurement to compare stabilization degree of videos.However,traditional methods of video stabilization assessment in Euclidean Space and Frequency Space are not consistent with the subjective evaluation,which affects the performance assessment and practical application of the video stabilization algorithm.In this paper,we have established two kinds of video stabilization quality assessment methods: full reference and no reference.These two methods firstly regard the motion path formed by the inter frame transformation of video as a curve in Lie Group space,and then quantitatively analyze the stabilization degree of video motion.Aiming at full reference video stabilization quality assessment,we propose a method based on geodesic distance of Lie Group space.Due to the full reference method needs real stable video as the benchmark,this paper designs a new method to construct data sets containing real stable video and corresponding jitter video to avoid the problem that the synthesized jitter is not real enough.When evaluating the stabilization degree of video,we map the motion paths of the real stable video and the evaluated video to the Lie Group space forming curves,and finally calculate the geodesic distance between the two curves,and take it as a measure to evaluate the stabilization quality of the video.Experimental results on the constructed data sets show that the correlation between the assessment result of the full reference method proposed in this paper and the subjective evaluation result is 95%.Aiming at no reference video stabilization quality assessment,we propose a method based on geodesic curvature of Lie Group space.The method uses discrete geodesic approximation to calculate the total curvature of motion path,and takes it as a measure to evaluate smoothness of the motion path,and then evaluates the stabilization degree of video.The experimental result shows that the correlation between the results obtained as a result of the proposed algorithm and the subjective evaluation results is 97%.Our algorithm is superior to the existing evaluation algorithms,such as inter frame fidelity and frequency domain analysis,which are improved by more than 20%.